Meta Muse Spark 1.3 pricing — the 21x discount is paid for with your prompts
Meta's Muse Spark 1.3, released in the first days of September 2026, sells the same model at two prices. The standard endpoint runs 1.25 dollars per million input tokens and 4.25 dollars per million output, the same as version 1.2, with cached input at 0.15 dollars, and Meta says data on this endpoint stays private. Alongside it sits a Contributor endpoint at 0.10 dollars input and 0.20 dollars output — 12.5 times cheaper on input and 21.25 times cheaper on output. The condition is a single clause: permission for Meta to use the developer's prompts and the model's completions for training. Paying with data instead of money is an old arrangement, but until now it has been buried in enterprise contracts or applied silently to free consumer tiers. Publishing it as a posted rate is the new part. On capability, the shipping xhigh configuration scores 61 on the Artificial Analysis Intelligence Index at 0.55 dollars per task, while the higher scores Meta cites come from a max configuration developers cannot broadly use yet
The three lines
- Price — standard 1.25/4.25 dollars per million tokens; Contributor 0.10/0.20. A 21x gap on output
- Condition — the cheap endpoint requires permission to train on your prompts and the model's replies
- Caveat — the best benchmark scores come from a max configuration still in safety testing; developers get xhigh
Key questions
- How much does Muse Spark 1.3 cost
- **Two prices, depending on the endpoint.** The standard endpoint is **1.25 dollars per million input tokens, 4.25 dollars per million output, and 0.15 dollars for cached input** — unchanged from version 1.2. The Contributor endpoint is **0.10 dollars input and 0.20 dollars output**. That is **12.5x** on input and **21.25x** on output. Which multiple you actually feel depends on your workload: most production systems send far more input than they receive back, so the effective saving sits nearer 12.5x, but output-heavy uses such as summarisation, translation and code generation move toward 21x. For context, the same week produced other pricing moves. Anthropic's Fable 5.1 cut cache-read pricing by 75 percent while leaving base rates alone, and Google's Gemini 3.8 Flash landed at 0.75 dollars per million tokens. **The Contributor rate of 0.10 dollars undercuts even that Flash-class price** — an unusual place for a model marketed on frontier performance.
- What happens to my data on the Contributor tier
- **Meta gains permission to train on it.** The published condition is permission to use prompts and completions for training, and three details matter. First, **the scope is not only what you send.** Model outputs are included, and outputs carry the shape of your inputs. Second, **your inputs contain your users' words.** If you build a product on the Contributor endpoint, the sentences typed by the people using that product travel with it, and the duty to tell them stays with you, not with Meta. Third, **the standard endpoint is described as keeping data private.** So this is not a default-on arrangement with a paid escape hatch. There are two doors and the developer picks one. For anything covered by privacy, medical, financial or client-confidentiality rules, the 21x discount is not a discount you are allowed to take. Note also that not training is a different promise from not retaining — providers commonly keep data for abuse monitoring or legal obligations regardless.
- Is the performance really frontier-class
- **The number Meta cites and the model you can buy are not the same configuration.** Meta reports frontier-level gains, particularly on long-running agent tasks. The shipping **xhigh** configuration scores **61 on the Artificial Analysis Intelligence Index** at **0.55 dollars per task**, described as the lowest cost per task among models at that intelligence level. That combination is genuinely competitive. The complication is above it. Meta's **max reasoning configuration** posts higher numbers on several benchmarks — 1,754 Elo on GDPval-AA v2 against 1,709 for xhigh — but it is **still completing additional safety testing and is not broadly available.** What developers can reach today through the Muse Code harness and the Meta Model API is xhigh. This is a recurring problem in model launches: a single model name covers several configurations that score differently, so a headline benchmark figure means little without the configuration attached to it.
Meta shipped Muse Spark 1.3 with two rate cards for the same model.
| Endpoint | Input / 1M tokens | Output / 1M tokens | Data condition |
|---|---|---|---|
| Standard | $1.25 | $4.25 | Kept private |
| Contributor | $0.10 | $0.20 | Prompts and completions used for training |
| Gap | 12.5x | 21.25x | — |
The standard prices are unchanged from version 1.2 (cached input stays at $0.15). The new thing is the second row.
1. The price did not change. The contract did.
Model prices have generally fallen in two ways. Either the same model is sold for less — Gemini 3.7 Flash halved overnight in August — or the provider bolts on a mechanism that spends less compute, as Anthropic did by cutting cache reads 75 percent. Both are the seller lowering its own costs.
The Contributor tier is neither. The compute cost is identical. The buyer pays part of the bill in data. The published condition is a single clause: permission to use prompts and completions for training.
The arrangement itself is old. Free consumer tiers have long been trained on, and large enterprise customers have long paid a premium for a no-training clause. What is new is that it now has a list price. A term that lived under the negotiating table has been printed on the menu.
2. Where 21x is available, and where it is not
Available — summarising or translating public documents, internal experiments, prototypes, open-source tooling. Anywhere the input holds no secret, the discount is simply a discount.
Not available — three cases:
- Anything carrying personal data. Your users' sentences become training material, and the duty to disclose that and obtain consent remains with whoever built the product.
- Health, finance, law. Sending client material into external training is frequently barred by regulation or contract before price enters the conversation.
- Internal work containing competitive information. Roadmaps, source code and unreleased results in a prompt are not a pricing question.
So the Contributor tier reads less as this is cheap and more as your input has been given a price. The 21x multiple is also the first public statement of what Meta thinks that data is worth.
3. The score and the product are different configurations
| Configuration | Intelligence Index | GDPval-AA v2 | Developer access |
|---|---|---|---|
| xhigh (shipping) | 61 | 1,709 Elo | Yes — Muse Code harness, Meta Model API |
| max reasoning | not published | 1,754 Elo | No — still in safety testing |
When Meta says "frontier performance," the strongest figures come from max, which is not broadly available. What developers can call today is xhigh.
xhigh is not a weak result on its own: 61 on the Intelligence Index at $0.55 per task, described as the lowest cost per task at that intelligence level. But the headline number and the purchasable product are not the same thing, and that gap is the reason a benchmark figure without a configuration name attached is close to meaningless.
4. What is still open
- The release date is inconsistent across sources. Trackers say September 2; VentureBeat describes a rollout starting September 3. Meta's own announcement time was not checked.
- The data terms are summarised, not published in full. Retention periods, whether deletion can be requested later, and how end-user disclosure is written into the contract were not verified.
- There is no date for max. Meta has not said what the remaining safety testing covers or when it ends. Worth reading alongside the fact that three of the four frontier models released in the first week of September shipped with gated cyber-capability tiers.
- Whether this becomes standard is unknown. If other labs copy it, "training consent" becomes a routine line on every AI rate card, and the data terms of cheap and free tiers become a far more direct negotiation than they are now.
Sources
- VentureBeat — Meta says Muse Spark 1.3 has frontier performance, but its best results come from a model developers can't broadly use yet
- Artificial Analysis — Muse Spark 1.3 Models: Intelligence, Performance & Price Comparison
- llm-stats — Muse Spark 1.3 API Pricing, Context Window & Benchmarks
- Codersera — Muse Spark 1.3: Pricing, Specs, and the Contributor Tier
- eesel AI — Meta Muse Spark 1.3: benchmarks, pricing, and what changed
- LLM Gateway — New AI Model Releases, September 2026 Timeline